Federated continual learning based on prototype learning
HaoDong ZHANG, Yang Liu, Jianliang Yu, Qinghua Hu, LiPing JING · Scientia Sinica Informationis · 2024
Federated learning allows multiple participants to collaborate on training models while preserving privacy. However, traditional federated learning methods do not support continuous learning and are not well-suited for dynamic scenarios. Recently, federated continual learning has emerged as a promising approach that enables ongoing learning and collaboration among participants. This scenario introduces additional complexities, such as catastrophic forgetting, heterogeneity, and limited communication resources. Considerably, this paper proposes a prototype-based federated continual learning approach. The proposed method uses prototypes for knowledge transfer, which improves communication efficiency and adaptability to model heterogeneity. Additionally, we introduce a mechanism to mitigate catastrophic forgetting through knowledge distillation and replay. We also provide a convergence analysis of the proposed method and validate its effectiveness through comparative and ablation experiments.